Enhance Part-Based Model for Person Re-Identification with Fused Multi-Scale Features

Abstract

In recent years, part-based models have been verified their effectiveness for person Re-identification (Re-ID). Since they learn an embedding only by partitioning single-scale features of the highest layer in the backbone network, their performances highly depend on the well-aligned parts of the extracted feature maps. However, misalignments occur very commonly in person Re-ID tasks due to the variations of viewpoints and poses. To address the part-misalignment problem and learn a more discriminative embedding for person Re-ID, we propose a novel Part-based model with fused Multi-Scale features (PMS), which innovatively upscales the low-layer features by using UpShuffle Modules and smoothly integrates the high-layer features. The fused multi-scale features are very robust to the variations of pedestrian scale and beneficial to resolve the part-misalignment problem. Experimental results on three commonly used datasets, including Market-1501, DukeMTMC-reID and CUHK03, have validated our model by outperforming the state-of-the-art methods with no need of re-ranking.